Self-attention–guided EfficientAttentionNet-B0 for robust and automated tuberculosis detection from chest X-ray images

Tuberculosis (TB) remains a major infectious disease burden, especially in resource-constrained environments where rapid and reliable diagnosis is critical for limiting transmission. Conventional chest X-ray–based diagnostic approaches are hindered by subjectivity in visual assessment and inconsistencies in image quality. While deep learning techniques have been introduced to automate TB screening, many existing architectures exhibit limited robustness and poor generalization across diverse datasets. To overcome these limitations, this study proposes a new architecture, EfficientAttentionNet-B0, which integrates EfficientNetB0 with a self-attention module to enhance TB detection from chest radiographs. The novelty of the proposed method lies in a task-specific, lightweight attention-guided EfficientNetB0 pipeline that combines transfer learning, self-attention, training-only oversampling, regularization, and 5-fold cross-validation for robust binary TB screening from CXR images. By employing self-attention, the network can prioritize the most informative regions within the images, thereby improving sensitivity to subtle radiographic manifestations of TB. The model was evaluated on a filtered benchmark dataset of 800 chest X-ray images using 5-fold cross-validation and achieved an average accuracy of 96.36%, along with high average values for precision, recall, and F1-score. Within the evaluated benchmark dataset and literature-based comparative setting, the proposed approach demonstrated competitive performance relative to the compared methods. Because the evaluation is limited to a single dataset, these findings should be interpreted as preliminary evidence of potential for automated TB screening support rather than evidence for immediate standalone clinical use.